across() improvements

[mgirlich]

Fixes the third point in #287.

• Refactor dt_squash_if() and dt_squash_across() to use across_setup(), very similar to dplyr::across_setup().
• The naming is now more consistent with dplyr. When passing a list to .fns missing names are not guessed from the function but filled with a number.
• Characters cannot be used as function anymore (like in dplyr).
• Fixed some minor issues for if_all().
• Increase test coverage for if_all().
• Use inject() instead of deprecated expr_interp()

[markfairbanks]

I like the simplification part of this PR - but before I do a code review I wanted to discuss the use of ... with a purrr style lambda.

Basically I think we should error when ... is used with a purrr style lambda (reasoning below).

In #287 @eutwt mentioned being able to do this:

library(dplyr, warn.conflicts = FALSE)

df <- tibble(x = 1:3, y = 1:3)

df %>%
  mutate(
    across(c(x, y), ~ .x + .y, 1)
  )
#> # A tibble: 3 × 2
#>       x     y
#>   <dbl> <dbl>
#> 1     2     2
#> 2     3     3
#> 3     4     4

However .y here could just as easily be ..2. And this can be extended to use more dots:

library(dplyr, warn.conflicts = FALSE)

df <- tibble(x = 1:3, y = 1:3)

df %>%
  mutate(
    across(c(x, y), ~ .x + ..2 + ..3 + ..4, 1, 2, 3)
  )
#> # A tibble: 3 × 2
#>       x     y
#>   <dbl> <dbl>
#> 1     7     7
#> 2     8     8
#> 3     9     9

@mgirlich maybe you have a good idea to deal with situations like this with an arbitrary number of dots? But instead of supporting .y (or the case of only one dot with a purrr style lambda), I would rather error if dots are used.

In the across() docs for dplyr 1.0.8 it mentions the use of dots is discouraged. So I think it makes sense to have them error in this edge case in dtplyr.

Open to thoughts though.

[eutwt]

Although I find the ability to use ... in across useful, I think I agree dtplyr probably should not support it given that it's likely to be deprecated in dplyr 1.1.0. Apologies for contradicting my own issue / not keeping that issue current.

[mgirlich]

Fine for me to remove to support for dots, especially for purrr style lambdas.
I didn't even know you could use ..2, ..3, etc for purrr style lambdas and to me your example just looks... terrible.

[markfairbanks]

I think we should definitely error in the case where dots are used with a purrr-style lambda. It doesn't currently work anyway so it won't break backwards compatibility.

However deprecating dots usage in general should be done with a deprecation lifecycle (if at all) since it currently works in dtplyr and also works in dplyr. I think a case like this should still work (without warnings) until dplyr makes the deprecation first.

df <- data.table(x = 1:3, y = 1:3)

df %>%
  summarize(
    across(c(x, y), mean, na.rm = TRUE)
  )

[markfairbanks]

Can drop rlang:: here

[markfairbanks]

Should this say something like "`dtplyr` does not support `...` in `across()` and `if_all()` when a lambda is used in `.fns`."?